[viff-devel] [issue75] Test without local computations

[Martin Geisler]

Good evening,

I've been doing some more performance measurements. This time I replaced
the shamir.share and shamir.recombine functions with fake versions (see
revision f7d1f7cd3dda). The replacements do a zero-degree sharing:

>>> viff.shamir.share(17, 1, 3)
[(1, 17), (2, 17), (3, 17)]

>>> viff.shamir.recombine([(1, 17), (2, 17), (3, 17)])
17

They skip the normal machinery and do:

  share     = lambda s, t, n: [(s.field(i+1), s) for i in range(n)]
  recombine = lambda s, x=0: s[0][1]

Enabling both fakes gives these timing results when doing 10000
multiplications on camel{10..12} on DAIMI (thyra02 is down so I used the
slightly slower camel machines):

    +-------------------+----------+----------+------+
    | Field size (bits) | Real (s) | Fake (s) | Diff |
    +-------------------+----------+----------+------+
    |         50        |   14.3   |   11.5   | 20%  |
    +-------------------+----------+----------+------+
    |        100        |   14.4   |   11.4   | 21%  |
    +-------------------+----------+----------+------+
    |        200        |   14.6   |   11.7   | 20%  |
    +-------------------+----------+----------+------+
    |        400        |   15.0   |   11.9   | 21%  |
    +-------------------+----------+----------+------+
    |        800        |   15.8   |   12.3   | 22%  |
    +-------------------+----------+----------+------+
    |       1600        |   18.4   |   14.0   | 24%  |
    +-------------------+----------+----------+------+

I think that is quite a saving!

Feel free to suggest other functions that I should replace with a fake
version. Ah, the PRSS functions would of course be an easy target... I
guess I'll take a look at them tomorrow.

-- 
Martin Geisler